What the project claims
The Apache 2.0 repository synchopate/llama-modchip, titled "CPU-First Inference for 284B & 30B Mixture-of-Experts Models on 16–64 GB RAM", was created on October 29, 2026 and last pushed on November 8, 2026 at 22:01 CET according to the GitHub API. The README for release v2.4.0 lists DeepSeek-V4 GGUF at 56.56 GiB (1.71 bpw), peak VmRSS 57.61 GiB at 256K context, and "0 GPUs required".
For a 64 GB high-memory workstation the same README table lists DeepSeek-V4-Flash 284B with "44.4 tok/s pp · 9.1–14.7 tok/s tg" and says it "Fits in 62.5 GiB usable Linux RAM (57.61 GiB VmRSS) with zero NVMe swap". A summary line also cites "44.36 tok/s 284B CPU prefill (16C Zen 5)" and a 30B MoE decode range of "25.0–111.6 tok/s" at 3.75 bpw. Total MoE parameters are given as 284.33B with 13.17B active.
What is not confirmed
These throughput and memory numbers come from the project README and its community leaderboard submissions. AppFor.ai did not run the software or re-measure tokens per second. The README does not, in the sections we read, publish quality scores for the 1.71 bpw DeepSeek-V4 quant. The project is community software that builds on llama.cpp-style tooling; it is not an official DeepSeek release.
What it means for you
If you already have a 64 GB CPU machine and want to try a heavily quantized DeepSeek-V4-Flash locally without a GPU, this repository is a community starting point to evaluate yourself. Treat the headline speeds as the author's measurements until you reproduce them, and check the license of both llama-modchip (Apache 2.0) and the DeepSeek weights you download.
Go to the source
Sources & further reading
Based on the official documentation below. Features and access can change; check the provider before making a decision.
Was this page useful?
Help us make our coverage more useful. This is feedback on the page, not a rating of the tool.
Go a little further